ConSel: Concept-Aware Self-supervised Learning for Regression Beyond Ordinal Tasks

Regression is a fundamental problem in computer vision, underpinning tasks such as gaze estimation, head pose prediction, age assessment, aesthetic quality evaluation, crowd counting and historical image dating. We introduce ConSel
(Concept-Aware Self-Supervised Regression), a unified framework that learns to predict continuous values by progressing from coarse semantic concepts to fine-grained numeric precision. ConSel follows a two-stage curriculum: (1) concept-aware self-supervised pretraining, which aligns visual embeddings with conceptual guidance through variance–covariance regularization without access to ground-truth labels, and (2) fine-tuning for precise continuous prediction. Unlike prior approaches that are optimized only for 1D ordinal regression, ConSel generalizes to both ordinal and multi-dimensional continuous tasks. Evaluated on 15 benchmark datasets spanning 6 domains, ConSel surpasses both domain-specialized and ordinal methods by 15-35% while using only 25% of labeled data (4× less than prior methods)}.